Downloaded 11/08/14 to Redistribution subject to SEG license or copyright; see Terms of Use at

Transcription

1 Velocity tomography based on turning waves and finite-frequency sensitivity kernels Xiao-Bi Xie*, University of California at Santa Cruz, Jan Pajchel, Statoil, Norway Summary We present a velocity tomography method, which is based on finite-frequency sensitivity kernels and applied to turning waves in models with generally positive velocity gradient. The finite-frequency sensitivity kernels are calculated using the one-way propagator. They replace the ray tracing to create links between the velocity model errors and the observed travel time residuals. The new approach is a wave-equation based method and wave phenomena are naturally incorporated. To demonstrate the inversion process, numerical examples are presented. For 2D velocity models with 0% to 30% velocity errors, we generate the synthetic data set. The travel time differences between seismograms from the true velocity model and those from the initial (intermediate) model are used as input data. The inversion is solved in an iterative way and the velocity models can be successfully updated. Introduction The turning wave data are widely used to determine the subsurface velocities, both at depth and near surface. Traditionally, the turning wave tomography has been dominated by the turning-ray method which assumes an infinitely high frequency. When a broadband signal propagates through a complex region, the ray theory often poorly approximates the actual wave phenomenon. Given this situation, a robust velocity tomography method based on the wave equation theory is preferred. Under this category, the full waveform inversion (FWI) has recently become very attractive. However, if the initial model is drastically biased from the true velocity model, the cycle skipping may cause convergence problem in the FWI. The travel time data are formed by accumulated multiple forward scatterings. As a substitution, we can use wave equation based sensitivity kernel for travel time tomography, which can tolerant large velocity errors and deal with data composed of large time delays. The sensitivity of finite-frequency signal to velocity perturbation has been extensively investigated and successfully used in velocity tomography (e.g., Woodward, 992; Vasco et al., 995; Dahlen et al., 2000; Huang et al., 2000; Zhao et al., 2000; Spetzler and Snieder, 2004; Jocker, et al., 2006; and Liu et al., 2009) and migration velocity analysis (e.g., Shen et al., 2003; Sava and Biondi, 2004; de Hoop, 2006; Xie and Yang, 2007, 2008a, b; Sava and Vlad, 2008; Fliedner et al., 2007; and Xia and Jin 203). Sensitivity kernels are usually derived based on the scattering theory. Several methods can be used to calculate finite-frequency sensitivity kernels in heterogeneous models. The method based on the full-wave finitedifference calculation (e.g., Zhao et al. 2005) can deal with complex velocity models but is very time consuming. On the other hand, the one-way wave equation based propagator (e.g., Xie and Yang 2007, 2008a, b) or the Gaussian Beam method (Xie 20) can be used to generate sensitivity kernels as well. The one-way method is efficient and can properly handle multiple forward scatterings, consisting with the travel time data. Given the large data sizes involved, this method is particularly useful in exploration seismology. We propose a turning wave tomography method based on the finite-frequency sensitivity kernels calculated by one-way wave propagators. Numerical examples are calculated to validate the proposed method. Finite-Frequency Sensitivity Kernel and its Calculation The finite-frequency sensitivity kernel can be defined as how the changes in velocity model will affect the observed finite-frequency seismic signals, e.g., their amplitudes, phases or travel times, or inversely, how the observed signals can sense the changes in velocity model. For a monotonic signal, its phase change can be linked to the velocity model perturbation as rs, r m r K Fr, rs, r dv. () V For a broad band signal, its travel time can be linked to the velocity perturbation as t( rs, rg) mrkbr, rs, r Gdv, (2) V where rs, r and t( rs, r G) are the phase and travel time differences for waves propagating from r S to r G and due to changes in model parameter mr vr v 0 r, is the frequency, v r is the velocity perturbation, v0 r is the background velocity. The spatial integral dv includes all regions with velocity perturbations. V The frequency domain phase sensitivity kernel K can be F expressed as (Woodward, 992; Spetzler and Snieder, 2004; Jocker et al., 2006) 2 Grr ; S, Grr ; KFrr, S, r imag 2k, (3) 0 GrG; rs, G r ; r, is the Green s function calculated in where 2 v0 r and from r to r 2, k0 v0 r, imag[ ] denotes taking the imaginary part. The broadband sensitivity kernel K for travel time can be obtained by stacking monotonic B kernels (Xie and Yang, 2008a) 756

2 W ( ) KBrr, S, rg KFrr, S, r d, (4) where W ( ) is a weighting function related to the source function (Xie and Yang, 2008a). Based on equations (3) and (4), calculating the sensitivity kernels is mostly calculating Green s functions in background velocity models. The oil and gas industry involves processing huge data sets and the efficiency is often a priority. The primary advantage of the one-way propagator is its high efficiency, often order of magnitude faster than the full-way finite difference method. The oneway propagator keeps most of the wave phenomena but neglects the back scattering and reverberations. It is suitable for simulating the travel time information generated by accumulated multiple forward scatterings. Figure. The stored kernel parameters FK. The enlarged parts show their details. The Inversion System Integral equation (2) forms the basis of velocity tomography, in which the sensitivity kernel K plays an B important role to associate the observed travel time residuals with the subsurface velocity errors between the true and trial velocity models. With observed travel time data t( rs, r G), we can invert velocity error vv 0 and use it in velocity model updating. Another important issue is the storage of sensitivity kernels. Unlike seismic rays, the finite-frequency sensitivity kernels are volumetric. Huge space is required to store thousands of kernels. This can cause problems in both storage and input/output. Several techniques can be used to reduce the storage space. Here, following Xie and Yang (2008b), we first partition the integral in equation (2) into the summation of integrals in small rectangular cells V r k t rs, rg m r K Br, rs, r G dv. (5) V r k k Within each cell, we use a hyperbolic function m r af r a2f2 r a3f3 r a4f4 r (6) to interpolate the unknown velocity perturbation, where f r, f r x, f r y, f r xy. (7) For the 4 corners of each cell, r 4, we have 4 equations, which link the coefficient a to the velocity values at the 4 i corners mr fr f2r f3r f4r a m 2 f 2 f2 2 f3 2 f 4 2 a r r r r r 2 (8) mr3 fr3 f2r3 f3r3 f4r3a 3 mr4 fr4 f2r4 f3r4 f4r4 a4 Substituting the values of hyperbolic functions at four corners, we have mr a m 2 x 0 0 a r 2, (9) mr3 0 y 0 a 3 m r4 x y xya4 where x and y are horizontal and vertical sizes of the cell. Using the inverse matrix x x 0 0 P, (0) y 0 y 0 xy xy xy xy equation (9) can express as ai Pm ij r j, () where repeating subscripts denote summation. Substituting (6) - () into the cell integral in (5), we have mrkbr, rs, rgdv V rk afk i i Br, rs, rgdv (2) V rk afk i ik Pm ij rj FKik where FK f r K r dv. (3) ik V r k i B Substituting equation (2) into equation (5) we have the linear system trs, rg Pm ij r jfk. (4) ik k j i 757

3 In this approach, we first determine the cell size according to the required accuracy for inversion. Within cell k, calculate integrals in equation (3) and store only 4 parameters FK for each cell. The accuracy of the kernel 4 storage is adaptive to the accuracy of the velocity model partitioning. Illustrated in Figure are parameters FK for about 000 kernels calculated for a velocity model, where selected kernels are enlarged to show their details. Equation (4) is the discretized version of the integral equation (2) and will be used in the following numerical tests. Figure 2. Snapshots of turning waves passing through velocity models with/without perturbations. Top: in the background model. Middle: in a model with -0% perturbations. Bottom: differential wavefield overlapped on differential velocity models. Numerical Examples () Turning waves in the perturbed velocity model. To illustrate how the velocity perturbations affect the turning wave propagation, shown in Figure 2 are snapshots of turning waves overlapped on the velocity models. These wavefields are calculated using the finite-difference method and a 5 Hz Ricker source. Shown in the top panel is the background model with a positive velocity gradient. In the middle panel, the model has a -0% perturbation which causes the wave front distortion and the amplitude focusing. Shown in the bottom panel is differential wavefield overlapped on differential velocity model. Note, even for such small perturbations, the wave front shift could be comparable to its wavelength. This may cause convergence difficulties for the conventional FWI. Figure 3. Top: synthetic seismograms in models with/without velocity perturbations. Middle: the velocity model with positive anomalies. Bottom: picked travel time differences from the synthetic data. Figure 4. Top panel: the background velocity which also serves as the initial model. Bottom panel: the true velocity model which has a positive velocity inclusion located at the center of the model. The middle panel is the inverted result after 4 iterations. 758

4 (2) Travel time data. To generate a synthetic data set, finite-difference method is used to calculate synthetic seismograms in the true velocity model and the initial (or intermediate) model. The travel time differences between two sets of seismograms are calculated using a cross correlation method. Illustrated in Figure 3 is such an example. The top panel is synthetic shot records in the perturbed and background models. Note the distance range where the travel times are affected by the velocity anomaly. Shown in the middle panel is the velocity model which has positive perturbations near the center part. The schematic wave path indicates the range the perturbations affecting the turning wave propagation. Shown in the bottom panel are picked travel time differences in this model. The data is composed of 35 shots each with 30 receivers. Each column is the result for a shot record. The patterns of the travel time delay provide the information for turning wave tomography. Figure 5. Travel time differences during iterations. (a) to (d) correspond to travel time differences before iterations to 4. Note different color bars are used for different panels. (3) Inversion example. In this example, the background is a -D model with a positive gradient and the true velocity model is the background plus a small patch with +0% velocity perturbations. The inversion is starting from the background model, which is shown on the top panel in Figure 4. The true velocity model is shown at the bottom. The inverted velocity model after 4 iterations is shown in the middle panel. Illustrated in Figure 5 are changes of travel time differences during the iteration process. The maximum travel time differences change from 0.5 s to 0.02 s. (4) Inversion example 2. In this example, we test a velocity model with large perturbations. In Figure 6, the top panel is the background velocity model which also serves as the initial model. The model in the bottom panel is the true velocity model. The maximum velocity perturbations are approximately 30%. Shown in the middle panel is the inverted velocity model after two iterations. The inverted result reproduces large-scale features of the original model but there are some small-scale artifacts due to the footprint of the cell size in the model partitioning. Certain smoothing during the iteration may solve this problem. Figure 6. Top panel: the background velocity model which has a vertical velocity gradient. This model also serves as the initial velocity model. Bottom panel: the true velocity model which is composed of the background velocity plus the velocity perturbation. Shown in the middle panel is the inverted velocity model. Conclusion We derived the theory for the turning-wave tomography based on finite-frequency sensitivity kernels. We developed the numerical method to calculate sensitivity kernels for the turning waves using the one-way propagator. This approach is highly efficient. A technique is designed for velocity model partitioning and efficiently store sensitivity kernels. We represent the unknown velocity model with interpolation functions and project the sensitivity kernels onto these functions. Instead of store the kernel itself, we store only its coefficients. With this technique, the storage is tremendously reduced, as well as the required I/O time. Because it is automatically adapted to the model partition, this method will not cause any accuracy loss. With the proposed inversion system, we successfully tested the turning wave tomography using synthetic data set. This method uses only the travel time information. Thus can handle initial models drastically biased from the true velocity model. The current results are preliminary. 759

5 References Dahlen, F., S. Huang, and G. Nolet, 2000, Fréchet kernels for finite-frequency traveltimes I, Theory: Geophysical Journal International, 4, de Hoop, M.V., R.D. van der Hilst, and P. Shen, 2006, Wave equation reflection tomography: annihilators and sensitivity kernels: Geophysical Journal International, 67, Fliedner, M.M., M.P. Brown, D. Bevc, and B. Biondi, 2007, Wave path tomography for subsalt velocity model building: 77th Annual International Meeting, SE Expanded Abstracts, Huang, S., F. Dahlen, and G. Nolet, 2000, Frechet kernels for finite-frequency travel times II, Examples: Geophys. J. Int. 4, Jocker, J., J. Spetzler, D. Smeulders, and J. Trampert, 2006, Validation of first-order diffraction theory for the traveltimes and amplitudes of propagating waves: Geophysics, 7, T67-T77. Liu, Y., L. Dong, Y. Wang, J. Zhu, and Z. Ma, 2009, Sensitivity kernels for seismic Fresnel volume tomography: Geophysics, 74, U35-U46. Sava, P.C., and B. Biondi, 2004, Wave-equation migration velocity analysis - I: Theory: Geophysical Prospecting, 52, Sava, P. and I. Vlad, 2008, Numeric implementation of wave-equation migration velocity analysis operators: Geophysics, 73, VE45-VE59. Shen, P., W. Symes, and C. Stolk, 2003, Differential semplance velocity analysis by wave-equation migration: 73rd Annual International Meeting, SE Expanded Abstracts, Spetzler, J. and R. Snieder, 2004, The Fresnel volume and transmitted waves: Geophysics, 69, Vasco, D.W., J.E. Peterson, Jr., and E.L. Majer, 995, Beyond ray tomography: Wavepaths and Fresnel volumes, Geophysics, 60, Woodward, M. J., 992, Wave-equation tomography: Geophysics, 57, Xia, F., and Jin S., 203, Angle-domain wave equation tomography using RTM image gathers, 83rd Annual International Meeting, SE Expanded Abstract, Xie, X.B., 20, Calculating finite-frequency sensitivity kernels using the Gaussian beam method, 8th Annual International Meeting, SE Expanded Abstracts 30, Xie, X.B., and H. Yang, 2007, A migration velocity updating method based on the shot index common image gather and finite-frequency sensitivity kernel: 77th Annual International Meeting, SE Expanded Abstracts, Xie, X.B., and H. Yang, 2008a, The finite-frequency sensitivity kernel for migration residual moveout and its applications in migration velocity analysis, Geophysics, 73, S24-S249. Xie, X.B., and H. Yang, 2008b, A wave-equation migration velocity analysis approach based on the finitefrequency sensitivity kernel, 78th Annual International Meeting, SE Expanded Abstracts, Zhao, L., T.H. Jordan, and C.H. Chapman, 2000, Threedimensional Fréchet differential kernels for seismic delay times: Geophysical Journal International, 4, Zhao, L., T.H. Jordan, K.B. Olsen, and P. Chen, 2005, Frechet kernels for imaging regional earth structure based on three-dimensional reference models: Bull. Seism. Soc. Am., 95, , doi: 0.785/